无人机多光谱影像的小麦倒伏信息多特征融合检测研究  

Multi-Feature Fusion Detection of Wheat Lodging Information Based on UAV Multispectral Images

作　　者：朱文静[1,2] 冯展康 戴世元 张平平 嵇文[4] 王爱臣 魏新华 ZHU Wen-jing;FENG Zhan-kang;DAI Shi-yuan;ZHANG Ping-ping;JI Wen;WANG Ai-chen;WEI Xin-hua(School of Agricultural Engineering,Jiangsu University,Zhenjiang 212013,China;Key Laboratory of Modern Agricultural Equipment and Technology,Ministry of Education,Jiangsu University,Zhenjiang 212013,China;Institute of Food Crops,Jiangsu Academy of Agricultural Sciences,Nanjing 210014,China;Institute of Agricultural Science and Technology,Jiangsu Hongzehu Farm Group Co.,Ltd.,Suqian 223900,China)

机构地区：[1]江苏大学农业工程学院,江苏镇江212013 [2]江苏大学现代农业装备与技术教育部重点实验室,江苏镇江212013 [3]江苏省农业科学院粮食作物研究所,江苏南京210014 [4]江苏省洪泽湖农场集团有限公司农业科技研究所,江苏宿迁223900

出　　处：《光谱学与光谱分析》2024年第1期197-206,共10页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(61901194);江苏省农业科技自主创新资金项目(CX(21)3061);国家自然科学基金项目(32001417);江苏省优势学科项目(PAPD-2018-87);江苏大学大学生创新训练计划项目(202110299634X)资助。

摘　　要：为探究多特征融合方法在作物倒伏领域快速精准识别中的适用性,利用无人机获取多田块冠层尺度的不同倒伏率麦田多光谱数据,对原始倒伏图像进行图像拼接、辐射校正、几何校正等预处理,并利用重归一化差值植被指数和阴影指数分别剔除土壤和阴影背景,提取小麦倒伏DSM模型和植被指数分别与多光谱图像进行多特征图像主成分变换融合,筛选差异性较大的纹理特征,采用支持向量机(SVM)、人工神经网络(ANN)和最大似然法(MLC)监督分类模型对多光谱和DSM融合图像、多光谱和归一化植被指数(NDVI)融合图像、多光谱图像和纹理特征图像进行监督分类,并采用总体精度(OA)、 Kappa系数和提取误差综合评价各监督模型的分类性能和倒伏提取精度。分类结果表明:各监督分类方法在不同倒伏区域提取结果建模效果趋势一致,SVM和ANN整体提取精度高于MLC,在高倒伏区域,多光谱与NDVI融合图像的SVM监督模型(OA:92.63%, Kappa系数:0.85,提取误差:1.11%)提取效果最好;在中倒伏区域,多光谱与DSM融合图像的SVM监督模型(OA:90.35%, Kappa系数:0.79,提取误差:9.34%)提取效果最好;在低倒伏区域,均值纹理特征图像的ANN监督模型(OA:91.05%, Kappa系数:0.82,提取误差:8.20%)提取结果较好。本研究将DSM模型、植被指数、纹理特征与多光谱图像进行融合对比,并对多特征融合方法能否高精度有效提取小麦倒伏信息进行了探究,结果表明无人机多光谱遥感结合特征融合技术能有效提取小麦倒伏面积,提取效果优于单特征小麦倒伏图像。本研究结果可为助力小麦倒伏灾情调查数据的精确获取方法提供参考。In order to explore the applicability of the multi-feature fusion method in the fast and accurate identification of crop lodging,this study used UAVs to obtain wheat field multispectral data with different lodging rates at multi field canopy scales.The original lodging image is preprocessed by image mosaic,radiometric correction,geometric correction,etc.,and the normalized difference vegetation index and shadow index are used to remove the soil and shadow background respectively.The wheat lodging DSM model and vegetation index were extracted and fused with the multispectral image for principal component transformation of the multi feature image,respectively,to screen the texture features with greater difference.Support Vector Machine(SVM),Artificial Neural Network(ANN)and Maximum Likelihood(MLC)supervised classification models are used to classify multispectral and DSM fusion images,multispectral and normalized vegetation index(NDVI)fusion images,multispectral images and texture feature images.The overall accuracy(OA),Kappa coefficient and extraction error were used to comprehensively evaluate each supervision model's classification performance and lodging extraction accuracy.The classification results show that the modeling effect of each supervised classification method in different lodging areas is consistent,and the overall extraction accuracy of SVM and ANN is higher than that of MLC.In the high lodging areas,the SVM supervised model(OA:92.63%,Kappa coefficient:0.85,extraction error:1.11%)of multispectral and NDVI fusion images has the best extraction effect;in the middle lodging area,the SVM supervision model(OA:90.35%,Kappa coefficient:0.79,extraction error:9.34%)of multispectral and DSM fusion images has the best extraction effect;in the low lodging area,the ANN supervised model(OA:91.05%,Kappa coefficient:0.82,extraction error:8.20%)of the mean texture feature image has a good extraction result.In this study,the DSM model,vegetation index,texture features and multi spectral images are fused and compared,

关 键 词：无人机遥感 图像处理 多光谱 特征融合 倒伏 小麦 

分 类 号：S127[农业科学—农业基础科学]